1. Field of the Invention
The present invention relates to a defect inspection method and apparatus which detects a difference between corresponding signals, compares the detected difference with a threshold value, and judges the part under inspection to be defective if the difference is larger than the threshold value. More particularly, the invention relates to an image defect inspection method and apparatus which detects a gray level difference between corresponding portions of two images, compares the detected gray level difference with a threshold value, and judges the part under inspection to be defective if the gray level difference is larger than the threshold value; the invention also relates to an appearance inspection apparatus which, by using such a method, detects a defect in a semiconductor circuit pattern formed on a semiconductor wafer. Still more particularly, the present invention relates to a technique for correcting the above threshold value in accordance with the signals (images).
2. Description of the Related Art
The present invention is directed to an image processing method and apparatus which compares corresponding portions between two images that should be the same, and judges the portion to be defective if the difference is large. The description herein is given by taking as an example an appearance inspection apparatus (inspection machine) for detecting defects in a semiconductor circuit pattern formed on a semiconductor wafer during a semiconductor manufacturing process, but the invention is not limited to this particular type of apparatus. Appearance inspection apparatuses generally used for such applications are bright field inspection apparatuses which illuminate the surface of a sample from a vertical direction and capture the image of its reflected light; however, dark field inspection apparatuses which do not directly capture the illumination light are also used. In the case of the dark field inspection apparatus, the surface of the sample is illuminated from an oblique or a vertical direction, and a sensor is disposed so as not to detect specularly reflected light. Then, the dark field image of the surface of the sample is obtained by sequentially scanning the surface with the illumination light. Accordingly, certain types of dark field apparatuses may not use image sensors, but it will be appreciated that the present invention is also applicable to such apparatuses. In this way, the present invention is applicable to any type of image processing method and apparatus as long as the method and apparatus are designed to compare corresponding portions between two images (signals) that should be the same and to judge the portion to be defective if the difference is large.
In the semiconductor manufacturing process, many chips (dies) are formed on a semiconductor wafer. Patterns are formed in multiple layers on each die. Each completed die is electrically tested using a prober and a tester, and any defective die is removed from the assembly process. In the semiconductor manufacturing process, manufacturing yield is a very important factor, and the result of the electrical testing is fed back to the manufacturing process and used for the management of each process step. However, as the semiconductor manufacturing process consists of many process steps, it takes a very long time before the electrical testing can be conducted after the manufacturing is started. Therefore, when, for example, a certain process step is found faulty as a result of the electrical testing, many wafers are already partway through the process and, thus, the result of the electrical testing cannot be easily utilized for improving the yield. In view of this, pattern defect inspection is performed to inspect formed patterns in the middle of the process in order to detect pattern defects. If the pattern defect inspection is performed at a plurality of steps in the manufacturing process, it becomes possible to detect any defects that occurred after the preceding inspection, and the result of the inspection can thus be promptly reflected in the process management.
In an appearance inspection apparatus currently in use, a semiconductor wafer is illuminated, an electrical image signal is generated by optically capturing an image of a semiconductor circuit pattern, and the electrical image signal is converted into a multi-valued digital signal (digital gray level signal). Then, a difference signal (gray level difference signal) is generated that represents the difference between the gray level signal of the pattern under inspection and the gray level signal of a reference pattern, and any portion where the difference is larger than a predetermined threshold value is judged to be a defect.
Generally, the reference pattern is a die adjacent to the die under inspection or a pattern of the same shape adjacent to the pattern under inspection. Then, defect sorting is performed in which each portion that has been judged to be defective is examined in further detail to determine whether the defect is a true defect that affects the yield. Defect sorting takes a long processing time because each defective portion needs to be examined in detail. Therefore, when judging defects, it is required that any true defect be judged to be a defect without fail, while minimizing the possibility of judging any other defect than a true defect to be a defect.
To achieve this, optimum setting of the threshold value is critical. If the threshold value is set too small, the number of pixels judged to be defective will increase, thus judging even portions not truly defective to be defective, and the resulting problem being that the time required for the defect sorting increases. Conversely, if the threshold value is set too large, even true defects may be judged not to be defects, resulting in the problem of insufficient inspection.
In a prior art method that automatically determines the threshold value according to the sample, the digital gray level signal of the pattern of a similar sample is generated in advance, followed by the generation of a gray level difference signal, and a histogram of gray level differences is constructed. Then, in the histogram, a variable reference difference is obtained which is set as a prescribed percentage of a portion where the gray level difference is large, and the threshold value for detection is determined by adding a fixed difference to the reference difference. This is to prevent the number of pixels judged to be defective from increasing appreciably in cases where the variance of the distribution of the differences is large, by considering that such cases can become a problem in practice. In this method, the variable reference difference varies from sample to sample, but the fixed difference to be added is fixed and does not vary from sample to sample; accordingly, this method has the problem that the proper threshold value cannot be determined when the noise level varies.
To solve this problem, various methods for determining the threshold value have been proposed. For example, Japanese Unexamined Patent Publication No. H04-107946 discloses a method that determines the threshold value based on the statistics of gray level differences computed at a plurality of portions of a pattern. More specifically, a histogram of maximum values is constructed by obtaining the maximum value of the gray level difference for each portion. Then, based on mean and standard deviations, the initial value of the optimum threshold value is set, and the optimum threshold value is determined by correcting the initial value based on the number of pixels detected as defective. This method, however, has the following problems: (1) the sample must be measured in advance and (2) inspection must be performed a plurality of times. Furthermore, while it is stated that the threshold value at which the number of detected defects abruptly changes is optimum, no description is provided of a specific method for obtaining such an optimum threshold value.
On the other hand, Japanese Patent No. 2996263 discloses a method that obtains an approximation curve from the relationship between the gray level difference and its frequency, and that takes the gray level difference at which the approximation curve becomes zero as the optimum threshold value. Here, the relationship between the gray level difference and its frequency is represented by a curve, but a curve does not necessarily become zero; therefore, there are cases where the approximation curve does not become zero. Further, even in the case of a straight line also, the straight line may not become zero, depending on its slope. Therefore, there can occur cases where the threshold value cannot be set. Furthermore, it is stated that the above curve can be obtained easily, but in actuality, this curve cannot be obtained easily because of its dependence on the distribution of gray level differences, and hence there arises the problem that the processing time increases.
Japanese Unexamined Patent Publication No. 2002-22421 discloses a method that performs a conversion to an error probability value by using a standard deviation. This method, however, involves the following problems: (1) as the standard deviation is computed directly from gray level differences, a large amount of computation is required and the processing time increases, and (2) as the error probability value, not the gray level difference, is used to judge the presence or absence of a defect, the error probability value must be computed for every gray level difference, and this again increases the processing time. There is the further problem that, because of the use of the standard deviation, the method is only applicable to a normal distribution and cannot be applied to other types of distribution.
For the inspection of semiconductor patterns, etc., it is desired to automate the inspection process, and it is also desired to automatically set the threshold value. To achieve this, the optimum threshold value must be set by instantly processing the detected gray level difference, and the presence or absence of a defect must be judged based on the threshold value; one possible solution would be to automatically set the threshold value by automatically performing a method such as described above. On the other hand, there is also a need to shorten the inspection time in order to improve throughput, but the above-described methods have such problems as the need to measure the sample a plurality of times in advance, the long processing time, etc. and therefore, are not suited for automating the threshold value setting process in a high-throughout inspection apparatus.
In particular, in the inspection of an actual semiconductor pattern, the noise level differs depending not only on the portion within a die but also on the position of the die on the wafer; furthermore, even when the same semiconductor pattern is formed, the noise level differs from one wafer to another. Therefore, it is required that the optimum threshold value be set by processing the gray level difference as it is detected, but none of the above-described prior art methods can satisfy such a requirement.
In view of the above background, the applicant of this patent application proposed the following image defect inspection method in Japanese Unexamined Patent Publication No. 2004-177397. That is, the distribution (histogram) of the gray level difference between corresponding portions of two images is constructed (see FIG. 1A), and its cumulative frequency is computed (see FIG. 1B). Then, assuming that the gray level difference has a distribution that obeys a prescribed type of distribution, a converted cumulative frequency is computed by converting the cumulative frequency so that the cumulative frequency shows a linear relationship to the gray level difference (see FIG. 1C). After that, an approximation straight line is computed by approximating the converted cumulative frequency by a straight line and, based on the thus computed approximation straight line, the threshold value is determined from a prescribed cumulative frequency value in accordance with a prescribed calculation method.
For example, in the example of FIG. 1C, the threshold value T is calculated as T=(P1−b+VOP)/a+HO, where “a” is the slope of the approximation straight line, “b” is the intercept at which the approximation straight line intersects the vertical axis, P1 is the cumulative frequency corresponding to a prescribed cumulative probability (p), and VOP and HO are prescribed sensitivity setting parameters.
As the converted cumulative frequency computed by this method shows a linear relationship to the gray level difference, subsequent processing for determining the threshold value is facilitated and, as a result, the threshold value can be set, automatically, in a short processing time.